This subproject builds on recent experimental and computational work in the Sander lab to further develop a systematic approach to the derivation of predictive computational models for cancer cells and tissues. This systems biology approach focuses on highly constrained (parameter-sparse) computational models derived from a series of systematic drug-drug perturbation experiments with rich observational readout. This approach (CoPIA - combinatorial perturbation-interaction analysis) is thus at an optimal methodological point between purely statistical network models on the one hand (that typically do not capture the physico-chemical dynamics of biological processes) and large differential equation models of biochemical kinetics (that typically lack sufficient parameters to realistically simulate cell biological processes). The project will start with detailed perturbation experiments and modeling in model systems for glioblastoma subtypes and for drug resistance in melanoma (synergy with sub-project 1). A key focus of the project will be in pushing forward the application of advanced measurement technologies and the development of novel, more efficient algorithmic approaches for network model construction ('reverse engineering') and optimization. Work on both of these tracks has already been initiated with one-year seed funding from the MSKCC Experimental Therapeutics Center (ETC). The Sander group has established a laboratory in the Zuckerman Research Center (2008) and implemented optimal growth conditions and drug drug perturbation protocols for both glioblastoma-derived tumorspheres and melanoma cell lines. Pilot experiments for high-throughput data acquisition using reverse phase protein arrays (RPPA, collaboration with Gordon Mills, MD Anderson) are in progress in June 2009. Algorithmic development has begun to focus on several promising routes. One challenge is to scale up model construction to systems with several hundred or more nodes, corresponding to a very large number of model configurations. We are currently exploring methods of statistical physics to make this process manageable, with the much more efficient derivation of probability distributions over discrete edge values rather than more expensive enumeration of sets particular solutions. As an alternative, we are exploring other algorithmic approaches, such as modular decomposition, starting solutions from correlation analysis and genetic optimization algorithms. In addition, we have designed an approach to incorporate secure prior knowledge into the modeling process. Continuous development of the computational approaches is a key thread running through all of the Specific Aims. The ultimate clinical aim of this sub-project is the design of effective combinatorial therapies that address diversity in tumor subtypes and differences in therapy-sensitive versus resistant tumors, validation of the combinatorial drug protocols in cell lines and mouse models, and the design of clinical trials.